The discussion of these different results extracted through the research questions defined, such as shown in Table 1, allows understanding the semantic interoperability scenario and the solutions usually applied in this scenario.
Conducting the search strategy
We chose seven different research bases to cover studies in the health and technology field, such as ACM Digital Library, IEEE Xplore Library, Science Direct, Springer Link. Moreover, we added Google Scholar to cover studies outside those bases. Our criterion was the relevance of these databases concerning the health and information technology literature. The search step in the databases, as mentioned above, aimed to index the search for studies published in the last ten years. Each database presents a way of formatting the survey, which we respect and modify to suit, but we kept all the mandatory terms defined in the PICOC strategy.
Finally, after applying the queries to the search bases, we had 6,032 articles. The initial filter aims to remove patents and citations, non-English studies, which resulted in around 783, roughly because some patents also appeared as citations.
As shown in Fig. 1, by year, the published articles in this area have been an interest constant in recent years. For this systematic review, the cutoff was September 22nd, 2020.
Article selection
Figure 2 shows the selection steps, removing impurity studies unrelated to the area of interest. Usually, these impurities studies had references related to the research area or citations of related works, however, without directly informing the area of interest. In addition, articles with less than six pages and no abstract, about 735 impurities, were excluded. We removed non-primary studies, such as editorials, chapters, thesis, reviews, and reports, approximately 1195 studies.
Then, we applied the inclusion and exclusion criteria based on the PICOC strategy: semantic interoperability, medical record (variants such as health record, medical record, patient record, hospital record), and standard. Finally, the inclusion and exclusion criteria are applied to the remaining studies to filter articles directly related to the research topic, totaling 2,288 excluded studies. The filters allow selecting studies according to defined terms.
Quality assessment
The last step had two parts; first, we filtered the studies by main interest area and evaluated the remaining corpus about its objectives. Many studies satisfied the inclusion and exclusion criteria but differed from the review’s interest topics. Thus, the quality assessment, the last step of the protocol, provides a cutoff parameter, where we look for studies published in relevant journals for the area of interest. Among these, we selected the highlighted studies. For this, we use the h5-index metric, which quantifies the relevance of newspapers and conferences in the last five years, a Google metric and works with the highest number H [24]. Some studies that presented relevant discussions but did not appear in this index were separated to contribute to our discussion section.
In the final selection, some bases showed a predominance of accepted articles. For example, ScienceDirect with eight studies followed for PubMed Medline with seven and Google Scholar with six studies. Next was SpringerLink with four and IEEE Explore with two, while ACM Digital Library and Web of Science had one from each research base. Figure 3 present the final corpus articles accepted in this systematic review.
Figure 3 shows the accepted articles distributed by year. Table 3 shows where each article was published (journal or congress), the number of articles per location, and the H5 index used in the quality assessment stage. Three times the journals had an above-average acceptance, a scenario explained by the applied area, which is related to the research question of this review. We highlight the International Journal of Medical Informatics, with five accepted studies, BMC Medical Informatics and Decision Making, and the Journal of the American Medical Informatics Association, with four accepted studies for each, a relevant number compared to the rest of the studies shown in Table 3 below.
Table 3 This graph presents the number of papers by year of publication after finishing the selection and filtering steps Data extraction
The information extracted from the articles aims to answer the question of interest in the review and identify approaches and solutions developed that allow achieving semantic interoperability in health records. Table 4 presents the final list of articles selected for this systematic review.
Table 4 This table presents the first author and respective studies by year, relating the publisher and the kind of place it was published Follows the analysis of the selected studies against the questions of interest, and each other answered individually.
SQ1 – What are the health standards adopted in the studies?
There is no consensus on a global standard for electronic health records, and the studies selected for this review reinforced this scenario. However, the extracted data shows a trend in the standard choice with a multilevel approach, as such openEHR, ISO/CEN 13,606, and HL7 formats. That dual model approach allows specialists in health and technology to perform in a joint work. Most of the studies had related advances toward a semantic dataset choosing standards to achieve semantic interoperability, as shown in Table 5.
Table 5 Health standards used in the selected studies The information extracted from studies demonstrates the open health standards as a trend, especially towards the two-level—openEHR and ISO 13606. Besides exchanging data and ensuring semantic interoperability concerns, as in [46], the authors had developed a framework combining ontology resources to predict high-risk situations in pregnancy. Additionally, the article contributed a summary analysis of 3 open standards, openEHR, ISO 13606, and HL7 CDA, showing their advantages and disadvantages. On the other hand, the studies [3, 4, 42] seek solutions for using openEHR and ISO13606 jointly, two open standards and similar definitions. This approach to approximation of the standards would allow the normalization of data. Although both standards use the ADL (Architecture Description Language), several differences in their types and definitions still need harmonizing.
In [16, 31, 43], the authors had explored slightly different opportunities from the known target of standards. For instance, [16] presented a methodology to represent the dependencies among data elements, concepts, and archetypes on a three-level Bayesian network and used the inference process to discover relevant archetypes—with promising results against the traditional search platform CKM. On the other hand, the authors in [43] improved patients’ current post-sale drug surveillance process. The data come from voluntary reports (spontaneous reports, yet just about adverse incidents). The EHR adoption would allow tracing a complete patient medical history and to predict potential risk factors.
The authors in [31] also explored a different opportunity. They developed a federated Metadata Registry/Repository (MDR) – a metadata database of data combining Common Data Elements (CDE) and HL7 CCD (Continuity of Care Document) models, proposing extensions to ISO 11179. Moreover, it was implemented in [36] the Health Level 7 (HL7) Virtual Medical Record (vMR) as a component service-based, that aims to collect patient data from different databases to allow the use in EHR data clinical decision support as a gateway between data sources and components.
SQ2 – What are the terminologies or health repositories used?
Terminologies and vocabularies can be understood as extensive collections of terms for a knowledge domain, making the language common. Terminologies aim to prevent local expressions, neologisms, and human typing from entering the EHR, thereby adopting formal classifications, e.g., diseases, events, procedures, specimens. Healthcare institutions share sensitive information, and systems must convey and not miss the meaning. One way possible is to build a local repository and manage an environment that applies proprietary concepts. However, the adoption of international terminologies ensures that using global terms and other classifications will have the same meaning on the other side – to any receiver. Table 6 present the terminologies usually adopted by the studies.
Table 6 The following are the international terminologies and classifications applied by the studies aiming at a shared vocabulary focusing on keeping the real meaning Table 7 The studies had more than one objective, so this table separates them into applications and proposed approaches to meet interoperability demands in health records, categorizing them according to the principal approach to developing the contribution Health standards generally easily adhere to different terminology as they are sets of data that can be entered into the repository and accessed and updated. Table 6 shows some terminologies most used by the selected studies. We highlight SNOMED-CT, the most cited terminology, bringing a clinically validated, semantically rich, controlled vocabulary. That facilitates evolutionary growth in expressivity to meet emerging requirements [50]. Also was often quoted and discussed, ICD is a diagnostic classification standard for clinical and research purposes. It defines a universe of diseases, disorders, injuries, and related health conditions [51], and in 2022 will publish the new version, the ICD-11.
While the adoption of standards is essential to achieve interoperability in EHR and effectively exchange information between different providers, the next level requires sharing knowledge and adding semantic value. A common language makes information comprehensible among who sent and received, allowing inferences in data and creating new connections from existing data. According to [49], semantic modeling essentially means linking words and terms to their senses, which is your main challenge.
Adopting a terminology engages more than one HIMSS level, as it condenses structural decisions regarding the system design, team, and organizational choices. The institution’s medical staff must accept adherence, use the terminologies, and publicly reflect that choice. After compliance around the organization and team, the information conforms to a standard and shares it with other providers.
SQ3 – What are the approaches used?
The selected studies reinforce the advantages of adherence to standards for achieving semantic interoperability. Exploring new approaches involving technologies and health standards to extract better results from consolidated standards has shown growth. Likewise, some studies using the use of semantic web technologies to meet information extraction and harmonization demands. Table 7 present the principal purpose of the selected studies.
Standards to allow semantic interoperability have been a common choice in electronic health records systems. The papers selected for this review strengthen that scenario and justify solving problems by developing solutions for this purpose. An initiative regarding making health records available for secondary use in [42, 43], also reinforces a consequent advantage of implementing semantic interoperability since the improved data quality is part of the process. These works are also strongly related to the application of representations of clinical models [15, 27, 42], OWL semantic structures, where they also serve as semantic mediators [15, 26, 31], whereas [30] also added an agent-based system to coordinate the community IHE.
Hundreds of biomedical ontologies are available in OWL format in the BioPortal repository, including many medical terminologies, which justified using ontologies in some studies to convert clinical models, data, and other terminologies to this representation format natural. According to the authors in [42], exploring data semantic representation in ontologies is justified because other structures usually have explicit connections between data, and ontological systems allow reasoning to build different meanings. For this reason, to represent EHR metadata into ontologies typically appear like a good choice for semantic goals.
Moreover, the use of ontologies has also proved promising for mapping scenarios of rules data access. As demonstrated in [30], an agent coordination infrastructure uses an OWL (Web Ontology Language) ontology to map the access rules of organizations from the community to the data. On the other hand, the authors in [28] developed an automatic extraction from semi-structured data using ontologies. They represented those concepts into ontologies of clinical vocabularies—another successful use of ontologies.
In line with the sharing and reuse of existing clinical models, some studies propose creating automated interfaces based on archetypes from the openEHR and ISO13606 standards, such as [29, 34]. On the other hand, some studies have identified novel architectures of service involving workflows in cloud environments, aiming to enable a tool to set resource pipelines, as presented in the works [3, 32, 36, 41], to access health data. Table 8 introduces the semantic web technologies used by the studies.
Table 8 The different semantic web technologies used in studies to solve semantic problems Despite being commonly used for semantic representation, semantic web technologies have broad applicability. Some studies have explored the representation of archetypes, rules, and relationships between different reference models (RM) of the openEHR and ISO136060 standards. The authors in [40, 44] have represented the reference model and archetype constraints into OWL ontology, which describes the instances and allows maintaining only one information (removing duplicated cases) to keep a relationship between RM and archetype.
Exploring other advantages, [3, 42] had presented EHR data into RDF and OWL structures – these transformations through the use of the Semantic Web Integration Tool (SWIT), as shown previously in Table 8. Additionally, they chose a graph database; respectively, the first used Neo4J Graph Database, and the last chose Virtuoso. However, both studies allow using Linked Open Data (LOD).
The studies also showed a trend by combining health standards and semantic web technologies to achieve semantic interoperability. There was no consensus about ontologies type adoption—OWL and RDF ontology. Both are used to represent the reference model and archetypes into ontologies. As reinforced by [4], the Archetype Definition Language (ADL) usually represents archetypes and has more of a syntactic orientation. Thus, it has disadvantages for achieving semantic interoperability, justifying the combining ontologies and clinical models. Also, the work presented in [27] used ontologies combining Semantic Web Rule Language (SWRL), and these rules aimed to allow applying logical deductive reasoning between archetypes and terminologies used.
This systematic review focuses on understanding the different approaches proposed to achieve semantic interoperability through health standards, and some studies have indicated promising results when selecting a specific database. Although there are unusual and justified databases in the studies as to the reason for selection, we also kept the most typical databases. Then, Table 9 present the database used by the studies.
Table 9 Databases used on studies The selected studies the not show a clear trend as to whether the health standard influences the choice of database, although it is possible to assess some characteristics. For example, the Virtuoso multi-model bank appeared in solutions using openEHR and HL7, not showing a dependency relationship with the standard type. However, a graph database allows ontologies querying if existent, reinforcing a possible hybrid solution with semantic web tools.
On the other hand, ARM and Think!EHR presented a framework designed to support archetype storage and allow querying in ADL. Therefore, there is a dependency of choice when using ISO 13606 and openEHR standards. The other storage structures did not show a strong relationship or dependence on health standards, and few studies discussed the type of database adopted by their solutions.
SQ4 – What are the main security concerns used?
One of the concerns sharing data is security, such as using anonymized data and methods for de-identification. Furthermore, the exchange between different organizations must encompass safe policies, even to exchange among proprietary systems — safety issues such as maintaining security and integrity without breaking the quality of patient data.
The studies used several kinds of data, with different sources, such as from patients coming from the real world, such as [46] pregnancy data, [3, 42] and [3] colorectal cancer screening protocols and lab tests, [43] pharmacoepidemiology data, [40]infants affected by Cerebral Palsy[45], coronary computed tomography angiography, [36] atrial fibrillation (AF), [33] diabetes mellitus, [37] chronic heart failure, [27] abnormal reactions and allergies patient’s data. Also, other studies used only results from lab tests as [26] radiology data, [35] histopathology reports, [38] test results of pertussis and salmonella. On the other hand, there was data synthetically generated in [30,31,32] that prevents any concerns about privacy or security.
The selected studies did not show a constant concern about data protection laws. This characteristic may be a consequence of the scientific character of the research since the application scenarios are generally academic environments and controlled. We highlight the two definitions observed regarding privacy and security concerns. In [32], HIPAA (Health Insurance Portability and Accountability Act) /HITECH and HITECH, and some directives from ISO 13606 applied on [3, 46], to store and manage health data.
On the other hand, we can consider that by keeping the security policies independent of the interoperability solutions, the authors aim to make the solutions generic in terms of local laws. Each country has its rules regarding using, sharing, and storing personal data – identification, financial or medical history. Thus, by keeping research an independent decision, studies are concerned with carrying out experiments that reflect real needs using real-world data.
Health Insurance Portability and Accountability Act (HIPAA) aims to ensure rights to citizens in the U.S., such as access to health records and request a copy of their data. As a patient, he can ask to correct something wrong with its health data, safe strategies to share data, and much more. To achieve that, they applied some methods to ensure security as authentication (SSO), authorization (OAuth), identity management, securing data at rest (using 256-bit advanced encryption standard), data in transit (HTTP with SSL), and auditing using a log data on the access. The last security layer represents the EHR stored alone because that provides data without the patient’s identity. Below, we highlight some characteristics regarding privacy, which some studies have indicated as necessary measures for using real-world data in their studies.
Studies that indicated the use of real-world data demonstrated some data providers’ concerns to enable the use of the information without compromising the patients. Therefore, we can observe that, regardless of the study, those who used real-world data need to apply some form of data anonymization. Table 10 present the two ways system providers usually allow data from real-world to academic researchers.
Table 10 The Table presents the usually followed guidelines inside academic and controlled environments to use data from the real world Table 11 Follows we list the evaluation approaches and the types of data used in the articles When the responsibility for data anonymization rests with the provider, the institution usually executes in a safe zone. Data only leave the health institution after being completely de-identified. On the other hand, the provider may allow researchers to access the system internally. In that case, the electronic registration system must have access levels for users, not allowing unrestricted access.
SQ5 – What are the evaluation approaches used?
The evaluation of an EHR involves storing data with quality, the end-users experience (health care professional), keeping complex information from specialist physicians, and ensuring the exchange of information between healthcare providers. Therefore, a scenario where the system must adapt to the routine of health professionals and facilitate data collection but meet the demands of sharing data with quality.
In this way, approaches to evaluation aim to identify measures to validate the different levels of interoperability (foundation, structural and semantic levels) that an EHR achieved and map the challenges of systematically doing it. The studies presented other evaluation methods, and Table 11 shows the different applied approaches.
Applying questionnaires as an evaluation method allows identifying which changes – according to the user and domain expert – would be more effective in improving the final solution through constant feedback. The author in [43] had an evaluation based on ISO/IEC 25,040 (SQuaRE), System Usability Scale (SUS), and Health IT Usability Evaluation Scale (Health-ITUES). On the other hand, functional tests allow the initial assessment to establish metrics to outcomes and automated tests.
However, none of the approaches assess the extent of interoperability or validate which different interoperability levels the applications have achieved – as per the levels defined by HIMSS [9]. Therefore, evaluating interoperability, as a result, is a challenge and an opportunity for further research, looking at more effective and unique methods.
SG1 – What is the state of art in health standards applied in health records?
There are some challenges to be faced by the health specialists and IT professionals to achieve semantic interoperability in EHR. This review identified different approaches to work around known problems and showed what technologies and strategies are being got in the area. Using the two-level (also called dual model) approach is not a global view of health standards. However, it can allow a common syntax and clinical data representation between the systems [38]. Furthermore, regardless of the management system adopted, this allows of making the clinical model software-independent: the specialized health professionals model the clinical document definition and terminologies as needed. In contrast, computer professionals are concerned with the structure, architecture, and choice of technologies to enable the use of knowledge defined by the health professional.
While there is no health standards consensus, we can see a trend in several studies using the two-level standard. This preference occurs because the OpenEHR standard provides a more mature public library – tools, archetypes, community support, and guides [34]. Furthermore, ISO 13606 is not publicly available, and developing novel archetypes to satisfy a system would take a lot of time [25]. Besides, to improve the adoption of the standards, many studies combined semantic web technologies and health standards with openEHR. A common practice identified in the articles was the structure representation of archetypes and reference models in ontologies, combined with rules mapping their constraints—another use involved ontologies representing metamodels with general information and ontologies for disease classification.
The solutions proposed by the selected articles brought different approaches to achieve semantic interoperability. In this scenario, Fig. 4 presents a taxonomy proposal around semantic interoperability in health records. Although the taxonomy is organized into five categories, the articles usually combine multiple themes to achieve the semantic interoperability challenge, so they are not exclusive sections. However, in the process of analyzing the studies, we defined five main categories: 1) Health Standards, 2) Classification and Terminologies, 3) Semantic Web, 4) Data Storage, and 5) Evaluation.
A taxonomy may allow us to assess a broader scenario—e.g., what artifacts involve building an EHR semantically interoperable. We reinforce that the development of an EHR must have semantic interoperability as a goal from early project design. The decision to adopt standards and terminologies at the beginning can facilitate the development process, once the artifacts are discussed as an integral part of the project, as mandatory.
The taxonomy scope involved the accepted articles, all technologies, standards applied to solve semantic interoperability and exchange data across organizations and health systems. The selected studies described many tools usually used to solve interoperability problems. However, we focus on solving the article research question and add that to the taxonomy list. We highlighted the taxonomy is not an exhaustive map to all semantic interoperability-related tools, only applied to the selected studies.
The first category, Health Standard, shows the standards used by the studies, with three subcategories. The dual model—openEHR and ISO 13606 inside share this architecture. Likewise, the HL7 standards have an ecosystem of standards, however, not all have the same goal. The HL7 organization also maintains a comprehensive list of terminologiesFootnote 3 compatible with available standards.
The choice of health standard may be related to some implementation team characteristics. The HL7 standard prioritizes a friendly relationship with the developer, with technical documentation and structures similar to the development ecosystem. In contrast, the openEHR proposes the opposite. Healthcare professionals have a user-friendly interface to create clinical models (in archetypes), focusing on defining knowledge. On the other hand, the artifacts defined for the openEHR standard—archetypes—have a structure that allows for in-depth detailing of clinical concepts, enabling interoperability at a semantic level.
Using standards inherently raises the concern of employing international terminologies to represent clinical terms and concepts. We show the vocabularies used in the Classification and Terminologies section of the taxonomy. Some terminologies have a massive adherence of studies, such as SNOMED and Logical Observation Identifiers, Names, and Codes (LOINC). Most health standards allow using more than one terminology in the same clinical document with semantic binding once a healthcare organization has legacy data and treats data through many vocabularies—enabled in openEHR and HL7, consequently in IHE.
Adopting terminologies as an inherent part of the EHR brings benefits, such as the standardization of everyday expressions, ensuring semantic contextualization concerning diseases, adverse events, and general classifications. The semantic contextualization allows new connections to collected data once the structure—patient’s history, lab tests, exams—has a semantic binding through international vocabularies. Unfortunately, there are still challenges when discussing the patients’ history. Usually have an open and unstructured text field, a format more accessible to physicians and health professionals. However, an unstructured text field does not readable to the machines.
The studies treat the unstructured data using semantic tools to extract and structure these data. Although there is a low variety of semantic web technology, the combination with health standards was almost unanimous. The most used tools often follow the definitions established by the W3C, such as OWL, RDF, SPARQL, SKOS, as reinforced in the recent research field review [52]. The Semantic Web section of the taxonomy shows the ontologies most used, such as SKOS, OWL, OWL-DL, and RDF as final semantic resources.
Although the studies do not explore the benefits of the choice of storage solution, some storage solutions are related to the type of health standard adopted. The trend towards the use of ontologies can impact the choice of storage solution since it is interesting to allow querying the ontological structure using SPARQL and exploring reasoning. In this scenario, Virtuoso, Neo4J, and Oracle databases present solutions that adequately meet ontologies’ storage demand.
In the Data Storage section, we map the different solutions used for storage to the taxonomy and categorize them according to the structure they provide. In addition, we highlight among these solutions two that specifically implement clinical document storage—archetypes such as ARM and Think!EHR. The other solutions use web-based graphical and semantic databases (W3C compliant), similar in their storage structure.
Finally, not all papers presented an evaluation format for their experiments or results obtained. Some authors had highlighted the use of case studies in a controlled environment to apply their final experiments. Often the authors used questionnaires with end-users to evaluate the user experience. The functional tests also had a prevalence, tracking specific modules with problematic scenarios while being developed to resolve before use. In some solutions, the authors described their partial evaluation methods during the development process. They wanted to evaluate tools’ accuracy or performance and not precisely the final solution, allowing for semantic interoperability. However, those methods did not consider the taxonomy because they did not influence the final discussion of the evaluated solution.
SG2 – What are the challenge and open questions to semantic interoperability in health records?
The adoption of standards has grown one of the ways to ensure semantic interoperability in health record systems. However, there are many barriers to overcome in the health organizations, such as legacy data, semi-structured data, non-structured textual data, complex systems that are sometimes not compatible for exchanging information. Once get over those challenges yet have internal adversities as the concepts, and medical terms used across the organization must preserve that meaning externally shared.
The search for interoperability at a semantic level involves combining different vocabularies from different areas to maintain a common meaning. For this scenario, the use of terminologies, ontologies, and global classifications plays an important role. However, since there is no worldwide consensus on health standards or unique clinical vocabularies, deciding which standard remains a technical issue when it should be an organizational one. Some research projects towards proposing vocabulary harmonization within EHR systems represent a promising alternative. These can allow the adoption of different vocabularies but with unique meanings. In other words, effective EHR communication depends on standardizing syntax, structure, and semantics (from the chosen architecture to the vocabulary used). The papers presented showed concerns about the difficulty in standardizing and normalizing data in legacy systems, as previously mentioned in [42, 43]. Complex models and legacy data are some of the limiters in secondary use and data reuse [42] and different purposes such as clinical research and decision support systems [38]. The lack of common terminologies and the extensive use of proprietary concepts inside EHR become the interoperability complex and requires a normalization process [43].